| Academic research is a process of continuous exploration and development.As the carrier of academic dissemination and exchange,publications carry important significance.Citation is an important task to confirm the proposition and concept.Citing proper citations can not only provide sufficient theoretical basis for papers,but also help readers to understand authors’ intention and ideas accurately.In the face of mass publications,citation recommendation provides an efficient and fast solution by automatically recommending appropriate citations for input text.Citation recommendation is classified into two categories,global and local,based on the differences in citation context.However,the existing local citation recommendation focuses on the academic research value,confusing the differences between the entity citation at word-level and the viewpoint citation at sentence-level in practical application,which leads to limited practical value.For this,this paper mainly completes two parts of work,one is that based on the usage scenario,the local citation is divided into two parts of word(entity)and sentence(viewpoint),and combined with the subjective needs of reviewers and objective statistical results,put forward the solution of entity reference to fill the gap of research.Another one is to optimize the existing citation recommendation methods on model and algorithm.Based on the needs of reviewers and writers,according to the unit granularity of the above tasks,this paper constructs a three-level citation recommendation model of words,sentences and papers,as a supplement and optimization of the existing classification.Specific research work and innovations are as follows:(1)Word level,entity citation,view of reviewers.Based on the statistical results of 25%of entity citation and the reviewer’s citation omission check requirement,this paper uses the historical citation records to build the mapping relationship between entities and articles,and the recommended citations are returned according to the weighted scores of semantic matching and citation count.The experimental results show that the proposed method is highly matched with entity references.(2)Sentence level,viewpoint citation,view of authors and reviewers.The location of citing sentences remains to be studied further in the existing research.In this paper,a discriminative model,based on the pre-training model SciBERT,is used to extract and learn the citation features of sentences for the problem of whether a sentence is a citing sentence or not.The comparative experiments show that the proposed discriminative model is superior to the previous method and can effectively identify and locate the citing sentences;(3)Paper level,overall citation,view of authors.This work introduces graph convolution network(GCN)into paper-level citation recommendation,and constructs an encoder-decoder model,which defines the node and edge in a different format compared with the previous work.Then graph embedding is used to encode the features of node and edge.Finally,the model returns the list of citations by decoding to reconstruct the graph.Experiments show that the GCN-based model works well in paper-level citation recommendation task. |
